The Wrong Practice

The library closed at eight, but the coding club room stayed lit until the janitor kicked them out. Maya had the laptop. Soren had the leaves.

Thirty-one photos. Fifteen healthy tomato leaves, sixteen sick ones with brown spots and curled edges, all shot against the same gray table so nothing would distract the program.

"Run it again," Maya said.

Soren pressed the key. The little training bar crawled across the screen and stopped.

"Fifty-two percent," he read. "Maya. A coin is fifty percent. We built a very slow coin."

"It's not enough pictures," she said. "It's learning from thirty-one leaves. You can't learn what a sick leaf is from thirty-one leaves."

"So we take more photos."

"We don't have more sick leaves. It's November."

Soren opened his notebook and wrote thirty-one, then drew a line under it and stared at the line. He didn't like the line.

"There's a thing," he said slowly. "In the tutorial we skipped. The boring one. You don't start from nothing. You download a network somebody already trained, and you build on top of it."

"Trained on what?"

He scrolled. "Cats. Dogs. Chairs. Bananas. A million pictures of stuff. It's the standard one everybody uses."

Maya made a face. "We don't have cats. We have tomatoes."

"I know."

"It has never seen a tomato leaf."

"I know. It's dumb. But the tutorial says do it, so." He shrugged. "Do the dumb thing?"

Maya was already reaching for the trackpad. "Do the dumb thing."

They loaded the cat-and-dog network. It was huge, hundreds of megabytes, all its knowledge already baked in from looking at a million animals and objects that had absolutely nothing to do with a sick tomato plant. Then they told it: forget the last part where you guess cat or dog. Keep everything else. Learn our leaves instead.

Soren pressed the key.

The bar crawled. Stopped.

"Read it," Maya said.

He didn't read it. He turned the laptop toward her.

Ninety-four percent.

Maya didn't say anything for a second, which was rare.

"Same thirty-one leaves," she said.

"Same thirty-one leaves."

"Nothing changed except we started from the cat one."

"Nothing changed except that."

Maya pushed back from the table. "That's wrong. That should not work. It knows dogs. Why would knowing dogs help it know a sick leaf."

Soren was already writing. His pen moved and then stopped in the middle of a word.

"Okay," he said. "Okay, what does a network actually learn first. When it looks at a dog."

"The dog."

"No. Not at first. At first it's just, there's a bunch of layers, right? The early layers are near the pixels." He tapped a photo of a healthy leaf. "The near-the-pixels part. What's actually in this picture, before it's a leaf."

Maya leaned in. "Edges. The edge of the leaf against the table."

"Edges. And?"

"Corners. Where the veins cross. Lines going different directions. Little patches of dark and light." She was talking faster. "The spot. The sick spot is just a round dark patch with an edge around it."

"And a dog," Soren said. "A dog is made of edges. Corners. Dark and light patches. Fur is a texture and a texture is just tiny edges over and over."

Maya stood all the way up now. "So the cat network never learned dogs first. It learned edges first. It had to learn edges to ever get to dogs."

"Right."

"And edges are edges." She pointed at the screen like it had confessed something. "An edge doesn't know if it's on a dog or a leaf. It's the same edge."

Soren stared at the leaf photo. He was seeing the million cat pictures underneath it now, invisible, holding it up.

"It spent a million pictures," he said, "learning how to see. The basic how-to-see part. Edges, corners, textures, curves. That part isn't about dogs at all. Anybody who learned to see anything would learn that first."

"So it's not starting over," Maya said. "It already knows how to look. We just handed it thirty-one leaves and said, point what you already know at these."

"And it did."

Maya sat back down slowly. "Soren. Real doctors do this. There are programs that read chest scans, right? Tumors and stuff. Somebody had to train those."

"Yeah."

"There aren't a million chest scans lying around. Sick people aren't a million cats." She looked at him. "Do they start those from the cat one too?"

Soren went quiet. He flipped back through his notebook to the thirty-one with the line under it, and above it he wrote: it learned to SEE on cats. it can look at anything now.

"They start from the cat one too," he said. "They have to. A network that already knows edges, pointed at a lung." He looked up. "The same edges that were on the dog are on the tumor."

Maya put both hands flat on the table and didn't move.

"That's the part," she said. "That's the part that's too big."

"What part."

"Learning one thing helps you with a completely different thing, as long as the underneath part is the same." She was almost whispering now. "Edges are underneath everything. So the practice counts. Even when it's practice on the wrong thing. Especially then."

Soren didn't say anything, because he was thinking about every wrong thing he had ever practiced. The notebook nobody else kept. The tutorials he read that turned out to be about something else.

The janitor's cart rattled in the hallway, coming closer.

"One more run," Maya said. "I want to see it again."

Soren loaded the cat network one more time, the whole enormous thing, a million animals folded up inside a program that was about to go looking at leaves.

He pressed the key. The bar began to crawl, and neither of them looked away from it.